Mapping sugarcane residue burnt areas in smallholder farming systems using machine learning approaches

Authors: PNVR, KOUTLYA - University Of Maryland; Bandaru, Varaprasad

Submitted to: International Journal of Applied Earth Observation and Geoinformation
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 10/22/2023
Publication Date: 10/23/2023
Citation: Pnvr, K., Bandaru, V. 2023. Mapping sugarcane residue burnt areas in smallholder farming systems using machine learning approaches. International Journal of Applied Earth Observation and Geoinformation. 6. Article 100347. https://doi.org/10.1016/j.atech.2023.100347.
DOI: https://doi.org/10.1016/j.atech.2023.100347

Interpretive Summary: Sugarcane crop residue burning is a common practice in sugarcane cultivated countries. Due to serious human health and environmental concerns, there have been efforts by countries to control sugarcane residue burning through placing regulation policies. Accurate mapping of burnt areas is an important step to assess the influence of regulations and improve existing policies to control burning. Satellite remote sensing has been widely used to develop methods to map burnt areas at large regional to global scales. However, these methods were reported to underperform to identify the crop residue burnt areas, particularly on smallholder farming systems. In this study, we evaluated two machine learning approaches (i.e., Support Vector Machines (SVM) and Artificial Neural Networks (ANN)) to map sugarcane burnt areas in a small-scale farming region in Thailand using Harmonized Landsat Sentinel-2 (HLS). Results suggested that both machine learning approaches performed well to identify the regional differences in sugarcane burning in the study region. The estimates from both ANN and SVM methods had significantly better accuracy compared to the operational MODIS global burnt area products. Between both machine learning approaches, the ANN based approach resulted in better estimates with higher accuracy compared to the SVM based method. Overall, our HLS satellite-based mapping algorithm with ANN classifier showed promise to monitor changes in sugarcane burnt areas reliably, and contribute to the successful implementation of existing policies to control residue burning in Thailand.

Technical Abstract: Satellite remote sensing methods have been proven to quantify the burnt areas resulting from different fire activities reliably. However, they were often reported to perform poorly to identify crop residue burnt areas, particularly in smallholder systems. This is primarily due to lack of frequent high spatial resolution satellite observations to determine the changes in seasonal crop specific characteristics required for burnt area mapping. In this study, we used Harmonized Landsat Sentinel-2 (HLS) satellite observations and evaluated two machine learning classifiers (i.e., Support Vector Machines (SVM) and Artificial Neural Networks (ANN)) to map post-fire sugarcane burnt areas for 2019-20 growing season in a dominant small-scale farming region in central and northeast Thailand. Results showed that both classifiers performed well to identify the typical spatial patterns of sugarcane burnt areas, even though ANN outperformed SVM, yielding higher accuracy at pixel and regional scales. At pixel level, ANN model accuracy was 93.4% while SVM’s best performed Polynomial kernel accuracy was 82.7%. The ANN estimated percent burnt area of 51.1% in the study region was closer to reported percent area (48.7%) by Thailand’s Office of Cane and Sugar Board (OCSB), compared to the SVM estimate of 62.9%. Relative to the estimated percent burnt area, total estimated burnt areas by ANN (315 thousand ha) and SVM (418 thousand ha) classifiers deviated more from OCSB reported area of 240 thousand ha. However, estimates from both classifiers had significantly better accuracy than the estimates of MODIS global burnt area products. Overall, this study demonstrated that HLS satellite observations provided spectral information required to build promising machine learning models to map burnt areas in smallholder farming systems with higher accuracy than global products. Our HLS satellite-based mapping algorithm with ANN classifier showed the potential to monitor changes in sugarcane burnt areas reliably, and contribute to the successful implementation of existing policies to control residue burning in Thailand.